MICA (Metrical Reconstruction of Human Faces)

The MICA dataset is a collection of eight smaller datasets unified under a common FLAME topology, covering approximately 2315 subjects. It provides registration meshes and corresponding fitted FLAME parameters.

Note: This dataset contains shape geometry only. To obtain the actual images for each subject, you must refer to the primary source datasets listed in the documentation.

The MICA dataset is for academic, non-commercial usage only.

Because MICA is an extension of existing datasets, the license is shared. You must also comply with the original licenses of each source dataset (e.g., Stirling, D3DFACS, Florence, LYHM, FaceWarehouse, FRGC, BP4D+), particularly regarding data privacy.

Accessing the FLAME results for the FRGC data requires a specific licensing process:

1. License the FRGC 2.0 dataset from CVRL via http://cvrl.nd.edu.
2. Once approved, you will receive an authorization email from Globus.
3. Use that authorization to email mica [AT] tue.mpg.de to request the MICA derivative work.

The easiest way to set up MICA is to use the provided install.sh script. This script handles the Conda environment creation and the downloading of required models. During execution, you will be prompted to provide your FLAME account credentials ({flame_user} and {flame_password}) to access the FLAME2020 model file server.

git clone https://github.com/Zielon/MICA.git
cd MICA
./install.sh

If you prefer to install manually, follow these steps:

1. Environment: Create the Conda environment using the provided file: conda env create -f environment.yml

2. FLAME2020 Model: Create an account at flame.is.tue.mpg.de, download the model, and place it in the /data/pretrained/FLAME2020/ directory.

3. MICA Pre-trained Model: Download the pre-trained model and place it in /data/pretrained/mica.tar.

4. InsightFace Models: Download antelopev2 and buffalo_l, unzip them, and place them in ~/.insightface/models/.

To train MICA, you need images from all eight datasets included in the MICA dataset collection.

If you are training from scratch and require Arcface model initialization, you must download Glint360K and specify its path in your configuration file using the key cfg.model.arcface_pretrained_model.

The dataset provides FLAME parameters stored in .npz files. You can load these using numpy and convert them to torch tensors for use in your models. The parameters include pose and betas.

import numpy as np
import torch

params = np.load('path.npz', allow_pickle=True)
pose = torch.tensor(params['pose']).float()
betas = torch.tensor(params['betas']).float()

flame = {
    'shape_params': betas[:300],
    'expression_params': betas[300:],
    'pose_params': torch.cat([pose[:3], pose[6:9]]),
}

To perform a quick demonstration of MICA, run the demo.py script. The script automatically processes all images located in the demo/input/ folder. For each subject processed, it generates:

• A .ply mesh file
• A rendered image
• A .npy file containing the FLAME parameters.

python demo.py

Each subset zip file follows this directory structure:

root\
    FLAME_parameters\
        actor_id\
            *.npz
    registrations\
        actor_id\
            *.obj